Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Still No Evidence for Cross-Fixational Integration: Non-Linear Encoding of Cortical Responses During Natural Vision

Carmen Amme¹, Philip Sulewski¹, Daniel Anthes², Malin Braatz¹, Peter König¹, Tim C Kietzmann¹; ¹Universität Osnabrück, ²University of Osnabrück

Presenter: Carmen Amme

Humans make several eye movements per second, requiring the visual system to construct perception from a sequence of brief fixations. This has led to the hypothesis that information gathered across fixations is integrated into a unified neural representation of the environment. In earlier work using linear encoding models, it was found that neural responses during a fixation were best predicted by the currently fixated image patch, with no measurable contribution from the previous fixation. Here, we tested whether such integration might instead be detectable through non-linear interactions. Using a large-scale MEG and eye tracking dataset from participants freely viewing natural scenes, we compared encoding models based on the current and previous fixation against matched control models in which the previous fixation was replaced by a random patch. Using non-linear MLP and GRU encoding models and examining responses across the whole cortex, we found no improvement in prediction performance when the true previous fixation was provided. Our findings therefore provide no evidence that visual information from the immediately preceding fixation contributes to neural responses during the current fixation, even when modeled with more expressive non-linear encoders.

Topic Area: Computational Models of Vision & Visual Cortex